Cluster Filesystem on FreeBSD

[Erik Scholtz, ArgonSoft GmbH]

Doug,

have you ever tried to put a NFS-volume into the working-path of an 
apache-server? I did, and I could messure a significant performance-loss 
of the webserver by 1sec per page-load.

The NFS was mounted on a 1GBit dedicated connection between the 
webserver and the nfs-server; I tried several TCP-options and 
MTU-Settings without any important change on the performance-loss.

Compared to a dedicated mounted ISCSI-Volume this 1sec loss is a lot! I 
think NFS is great for changing big amounts of data. But for short 
read/write-access NFS does not seem to be the first choice.

Greetings,
Erik
-- 

My blog: http://blog.elitecoderz.net



Doug Poland wrote:
> On Mon, March 1, 2010 12:11, Leinier Cruz Salfran wrote:
>> On Sun, Feb 28, 2010 at 6:09 AM, Erik Scholtz, ArgonSoft GmbH
>> <escholtz at argonsoft.de> wrote:
>>
>>> I did some research the last two weeks on how to build a cluster
>>> filesystem on FreeBSD.
>>>
>>> My solution at the moment is, to rsync all filesystems once a
>>> minute, which is rather to rare. So I tried to get a hook with
>>> KQueue to rsync the filesystems on data-change. Unfortunatly I could
>>> not find a working solution (had a try with IO::KQueue using perl).
>>>
>> i use rsync to make partial data backup .. ie: /etc, /usr/local/etc,
>> /usr/home, /var/logs ...
>>
>>> How do you guys solve this problem (of a shared filesystem with
>>> rw-option)?
>>> Any hints are welcome, since I'm getting very frustrated at the
>>> moment.
>>>
>> there is a project named 'hast'[1] for a clustered filesystem .. it's
>> being developed by pawel .. the project has some completed milestones,
>> so you can get it from fbsd src svn tree .. hast can do clustered
>> filesystem right now but it's not complete, so there is no stable yet
>>
>> other way is gmirror[2] + ggated .. with that you can get a raid1 over
>> net solution .. but i think it's not prepared to be used as
>> master-master soluction
>>
> Neither hast nor gmirror+ggatd are cluster filesystems, in that only
> one "side" of the storage is available for writes at a point in time. 
> Filesystems like OCFS2 and GFS allow multiple, simultaneous read-write
> access to block devices.
> 
> Given there is not true cluster filesystem available for FreeBSD at
> this time, I wonder aloud why so many people are so quick to dismiss
> NFS?  NFS provides "most" of features of a cluster filesystem today.
> If one were to choose NFS for shared storage, one could use tools
> available today to make NFS highly available (hast, gmirror+ggated).
> 
> 
>